Do Knowledge Graph Embedding Models Capture Entity Similarity as Intended?

The core message of this paper is that the widespread assumption that knowledge graph embedding models (KGEMs) create semantically meaningful representations of entities by positioning similar entities closer in the embedding space does not hold universally. The authors show that different KGEMs exhibit varying degrees of adherence to this "KGE entity similarity assumption", and that performance in link prediction tasks does not reliably correlate with the ability to group similar entities together.